Fix notebook Part 3 colorbar overlap and Part 4 ELECTRA examples

Part 3: Place colorbar in dedicated axes to prevent overlay on right
heatmap. Part 4: Replace chef cooked/ate examples (too similar) with
doctor/musician context that triggers fake detection. Update
visualization sentence to more plausible research context.

Author: jeremymanning

slides/week6/bert_variants_demo.ipynb

@@ -272,19 +272,20 @@
         "    ax.set_yticklabels(labels, fontsize=7)\n",
         "    return im\n",
         "\n",
-        "fig, axes = plt.subplots(1, 2, figsize=(16, 7))\n",
+        "fig, axes = plt.subplots(1, 2, figsize=(14, 6))\n",
         "\n",
         "print(\"Computing embeddings for BERT...\")\n",
         "bert_embs = get_embeddings(\"bert-base-uncased\", sentences, pooling='cls')\n",
-        "plot_similarity(bert_embs, \"BERT [CLS] Similarity\", axes[0])\n",
+        "im1 = plot_similarity(bert_embs, \"BERT [CLS] Similarity\", axes[0])\n",
         "\n",
         "print(\"Computing embeddings for RoBERTa...\")\n",
         "roberta_embs = get_embeddings(\"roberta-base\", sentences, pooling='mean')\n",
-        "im = plot_similarity(roberta_embs, \"RoBERTa Mean-Pooled Similarity\", axes[1])\n",
+        "im2 = plot_similarity(roberta_embs, \"RoBERTa Mean-Pooled Similarity\", axes[1])\n",
         "\n",
-        "fig.colorbar(im, ax=axes, shrink=0.8)\n",
         "plt.suptitle(\"Semantic Clustering Across Variants\", fontsize=16, color='#00693e')\n",
-        "fig.tight_layout(rect=[0, 0, 1, 0.95])\n",
+        "fig.tight_layout(rect=[0, 0, 0.88, 0.95])\n",
+        "cbar_ax = fig.add_axes([0.90, 0.15, 0.02, 0.7])\n",
+        "fig.colorbar(im2, cax=cbar_ax)\n",
         "plt.show()"
       ]
     },
@@ -338,13 +339,18 @@
         "        status = \"FAKE\" if prob > 0.5 else \"REAL\"\n",
         "        print(f\"{token:<12} | {prob:<10.4f} | {status}\")\n",
         "\n",
-        "# Example 1: A natural sentence\n",
-        "detect_fake_tokens(\"The chef cooked a delicious meal for the guests.\")\n",
+        "# Example 1: A natural sentence (all tokens should be marked REAL)\n",
+        "detect_fake_tokens(\"The doctor examined the patient carefully.\")\n",
         "\n",
         "print(\"\\n\" + \"=\"*40 + \"\\n\")\n",
         "\n",
-        "# Example 2: A sentence with a 'fake' token (replaced 'cooked' with 'ate')\n",
-        "detect_fake_tokens(\"The chef ate a delicious meal for the guests.\")"
+        "# Example 2: Same sentence but 'examined' → 'watched' (plausible but wrong in context)\n",
+        "detect_fake_tokens(\"The doctor watched the patient carefully.\")\n",
+        "\n",
+        "print(\"\\n\" + \"=\"*40 + \"\\n\")\n",
+        "\n",
+        "# Example 3: 'doctor' → 'musician' (semantically odd with 'patient')\n",
+        "detect_fake_tokens(\"The musician examined the patient carefully.\")"
       ]
     },
     {
@@ -354,21 +360,24 @@
       "outputs": [],
       "source": [
         "# Let's visualize the discriminator's confidence\n",
-        "sentence = \"The computer programmed the human to write better code.\"\n",
+        "# 'published' has been swapped in for 'presented' — plausible but detectable\n",
+        "sentence = \"The researcher published her findings at the annual conference.\"\n",
         "inputs = tokenizer(sentence, return_tensors=\"pt\")\n",
         "tokens = tokenizer.convert_ids_to_tokens(inputs[\"input_ids\"][0])[1:-1] # Remove CLS/SEP\n",
         "\n",
         "with torch.no_grad():\n",
         "    logits = model(**inputs).logits[0][1:-1]\n",
         "    probs = torch.sigmoid(logits).numpy()\n",
         "\n",
-        "plt.figure(figsize=(10, 5))\n",
+        "plt.figure(figsize=(12, 5))\n",
         "colors = ['#9d162e' if p > 0.5 else '#00693e' for p in probs] # Red for fake, Green for real\n",
         "plt.bar(tokens, probs, color=colors)\n",
-        "plt.axhline(y=0.5, color='gray', linestyle='--')\n",
+        "plt.axhline(y=0.5, color='gray', linestyle='--', label='Decision boundary')\n",
         "plt.ylabel(\"Probability of being 'FAKE'\", color='#9d162e')\n",
         "plt.title(\"ELECTRA Discriminator: Identifying 'Fake' Tokens\", fontsize=14)\n",
         "plt.ylim(0, 1)\n",
+        "plt.legend()\n",
+        "plt.tight_layout()\n",
         "plt.show()"
       ]
     },
@@ -378,10 +387,10 @@
       "source": [
         "### \ud83d\udca1 Discussion\n",
         "\n",
-        "- In the second example, did ELECTRA correctly identify \"ate\" as the fake token? Why might it be suspicious of that word in that context?\n",
-        "- How is this task different from BERT's masked language modeling? Why might it be more efficient?\n",
-        "- Try a sentence where you replace a word with a synonym. Does ELECTRA still flag it as fake?\n",
-        "- What happens if you give it a completely nonsensical sentence?"
+        "- In Example 2, did ELECTRA flag \"watched\" as suspicious? Doctors *examine* patients — \"watched\" is grammatical but semantically odd in a clinical context.\n",
+        "- In Example 3, did swapping \"doctor\" → \"musician\" get detected? Why might that be easier or harder to detect than swapping the verb?\n",
+        "- How is this task different from BERT's masked language modeling? Why might it be more efficient? (Hint: how many tokens does each approach learn from per sentence?)\n",
+        "- Try replacing a word with a close synonym (e.g., \"quickly\" → \"rapidly\"). Does ELECTRA flag it?"
       ]
     },
     {